This experiment using the public domain OpenMIIR dataset has been described in the paper Sebastian Stober: Learning Discriminative Features from Electroencephalography Recordings by Encoding Similarity Constraints. In: Proceedings of 42nd IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP'17), 2017.
Please cite this paper if you use any of this code!
Note that this is a revised and extended version of an experiment originally described in Sebastian Stober; Avital Sternin; Adrian M. Owen & Jessica A. Grahn: Deep Feature Learning for EEG Recordings. In: arXiv preprint arXiv:1511.04306 2015.
This code heavily depends on Theano, Blocks and Fuel. Using CUDA/cuDNN is optional but strongly encouraged. Further dependencies comprise MNE-Python for pre-processing and plotting, librosa for pre-processing, as well as "usual suspects" like numpy, scikit-learn, joblib etc. Make sure these libraries are installed properly if you want to run this code!
-
data/
Pre-processed data files. See separate README! -
deepthought/
Deep-learning-related code. Refactored and extended version of the legacy deepthought code that was based on the discontinued Pylearn2 framework. -
mneext/
Alternative resample method for MNE-Python, copied from the legacy deepthought code. -
openmiir/
Code specific to the OpenMIIR dataset, adapted from the legacy deepthought code. -
results/
Pre-computed network parameters and results for the "Train..." jupyter notebooks. Directory names match with the job_id given in each notebook.
This research was supported by the donation of a Geforce GTX Titan X graphics card from the NVIDIA Corporation.
Sebastian Stober <sstober AT uni-potsdam DOT de>
Research Focus Cognitive Sciences
Machine Learning in Cognitive Science Lab
University of Potsdam
Potsdam, Germany